Infant Movement Classification Using Pose-Based Motion Features and a Learned Hybrid k-Nearest Neighbor Model

Background/Objectives: Objective assessment of infant movement remains challenging because clinical assessments mostly depend on expert visual inspection and subjective scoring. This study introduces a pose-based movement analysis framework that converts raw RGB infant videos into resilient spatiotemporal representations and executes multi-class infant movement classification utilizing a lightweight learned hybrid k-nearest neighbor (kNN) model. Methods: The proposed pipeline extracts 2D body keypoints from each video frame, applies shoulder-centered spatial normalization to reduce scale and translation variability, and divides recordings into overlapping temporal windows. Within each segment, a small set of interpretable kinematic descriptors captures joint speed statistics, bilateral symmetry, inter-limb coordination, postural distances, movement smoothness, directional variability, entropy, and ratio-based relational features. This results in a fixed 60-dimensional feature vector for each segment. The pose-derived features were evaluated using a learned hybrid kNN classifier with adaptive similarity weighting and conventional classifiers under the same subject-wise evaluation protocol. The framework was evaluated using both segment-wise stratified 5-fold cross-validation and subject-wise grouped 5-fold cross-validation. Results: Segment-wise CV yielded an accuracy of 96.44%, while subject-wise grouped CV achieved 94.80%, indicating strong performance under both evaluation settings. Conclusions: The study contributes a pose-based framework and the BERA-GAZI-RGB dataset for infant movement classification. Learned Hybrid kNN performed similarly to conventional kNN, without demonstrating a consistent advantage of hybrid weighting. Further validation is needed to assess its potential for early movement screening.